An obvious (if naive) PDL approach would be to read images using PDL::IO::Pic, then multiply the values by a mask that was 1 for the transparent part, or (e.g.) 0.9 for the watermark part.